mindspore.nn.L1Regularizer

class mindspore.nn.L1Regularizer(scale)[source]

Applies l1 regularization to weights.

l1 regularization makes weights sparsity

\[\text{loss}=\lambda * \text{reduce_sum}(\text{abs}(\omega))\]

Note

scale(regularization factor) should be a number which greater than 0

Parameters

scale (int, float) – l1 regularization factor which greater than 0.

Inputs:

  • weights (Tensor) - The input of L1Regularizer with data type of float16 or float32. The shape is \((N,*)\) where \(*\) means, any number of additional dimensions.

Outputs:

Tensor, which dtype is higher precision data type between mindspore.float32 and weights dtype, and Tensor shape is ()

Raises

  • TypeError – If scale is neither an int nor float.

  • ValueError – If scale is not greater than 0.

  • ValueError – If scale is math.inf or math.nan.

Supported Platforms:

Ascend GPU CPU

Examples

>>> scale = 0.5
>>> net = nn.L1Regularizer(scale)
>>> weights = Tensor(np.array([[1.0, -2.0], [-3.0, 4.0]]).astype(np.float32))
>>> output = net(weights)
>>> print(output.asnumpy())
5.0